Tests as Instructions: A Test-Driven-Development Benchmark for LLM Code Generation

Yi Cui

Keywords: test, llm, coding, benchnark, instruction

Abstract:

This paper focuses on test-driven development (TDD) tasks, where test cases act as both instruction and verification for LLM code generation. We build a TDD benchmark to evaluate frontier models, where reasoning models of OpenAI achieve SOTA. We identify instruction following and in-context learning as the critical abilities for all models to succeed at TDD tasks. We further reveal their vulnerabilities to long instructions as an area of improvement.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8969